Media Intelligence

Making computers listen to desired sounds anywhere

- Probabilistic modeling and integration for speech enhancement -

Abstract

Sounds recorded using distant microphones are usually distorted by noise and reverberation, which degrades both sound quality and speech recognition accuracy. Conventional techniques for signal enhancement (SE) are limited to specific environments. To realize SE techniques that apply to arbitrary environments, we need to develop a universal model that probabilistically integrates models for specific environments. We have developed an SE technique that applies to a wider range of environments by integrating models for reverberant environments and diffuse noise environments. This research can be applied to, e.g., distant voice control of home appliances.

Photos

Poster


Please click the thumbnail image to open the full-size PDF file.

Map

Presentor

Nobutaka Ito
Nobutaka Ito
Media Information Laboratory
Keisuke Kinoshita 
Keisuke Kinoshita 
Media Information Laboratory
Takuya Yoshioka
Takuya Yoshioka
Media Information Laboratory
Masakiyo Fujimoto
Masakiyo Fujimoto
Media Information Laboratory
Marc Delcroix
Marc Delcroix
Media Information Laboratory
Araki Shoko
Araki Shoko
Media Information Laboratory
Espi Miquel
Espi Miquel
Media Information Laboratory
Tomohiro Nakatani
Tomohiro Nakatani
Media Information Laboratory